Enable auto optimization and improve classification
Use the procedure in this topic to enable auto optimization to improve classification.
Use the Auto Optimization window to optimize the parameters for the Adaptive Feature Classifier. The best practice is to use a training set for auto optimization. This is because the Classification Set may not contain enough documents to provide successful optimization results.
In many cases, the default settings are sufficient. If you want to modify these settings, the best practice is to generate a classification benchmark before optimization, and then another benchmark after each change. Recording all changes made along the way will ensure you can replicate your results.
Before you can enable auto optimization for an Adaptive Feature Classifier to improve classification, verify that you have performed the following:
- Open the Documents window if it is not already open.
- Open or select a document set and its document subset that contains documents for optimization.
-
On the
Process tab, in the
Train group, select
Optimize Content Classifier.
The Adaptive Feature Classifier - Auto Optimization window is displayed.
-
Optional. Before performing automatic optimization, select
Advanced to tweak the classifier settings.
The Optimization Settings window is displayed.
-
In the
Optimization Settings window, configure the following settings.
Setting
Description
Max. number of features
With respect to optimization, a feature is a string that is unique to a class. Trigrams (3-letters) or words are useful for helping with classification, but they are not unique enough, especially if you are using a fuzzy string match. Features are strong stings. For example, "Invoice" might be found in all classes, but "Invoice statement" is unique to one class and "This is the final statement" is unique to another class. (Default: Selected)
The feature length includes spaces. The maximum length is set to 7, this includes any spaces between words.
Min. values (Default: 1000) | Max. values: (Default: 10,000) | Step width: (Default: 1000).
Min. feature length
The minimum feature length determines the shortest length of a feature or string. Configures a minimum value and a maximum value as well as a step width for the minimum feature length. (Default: Selected)
Min. values: (Default: 3) | Max. values: (Default: 5) | Step width: (Default: 1)
When optimization is performed, the minimum feature length is tested for features that are 3, 4, and 5 characters long as a minimum.
Max. feature length
The maximum feature length determines longest length of a feature or string. Configures a minimum value and a maximum value as well as a step width for the maximum feature length. (Default: Selected)
Min. values: (Default: 16) | Max. values: (Default: 64) | Step width(Default: 8)
When optimization is performed, the maximum feature length is tested for features that are 16, 24, 32, 40, 48, 56, and 64 characters long as a maximum.
Min. feature frequency
Specifies how many times a feature is present in order for a document to be classified. (Default: Selected)
Min. values: (Default: 2) | Max. values: (Default: 5) | Step width: (Default: 1)
When optimization is performed, there must be at least 2 instances of a feature in order for a document to be classified. There can be 2, 3, 4, and 5 instances using the default values.
Min. class entropy
Entropy is the level of uniqueness in a project. A higher the value indicates the features are unique. The more classes that a project has means more entropy. (Default: Selected)
Min. values: (Default: 0.2) | Max. values: (Default: 0.8) | Step width: (Default: 0.1)
Values 0.2, 0.3, 0.4, 0.5 0.6, 0.7, and 0.8 are tested.
Use words only
Uses full words only. (Default: Selected)
Only full words are recognized. There are no partial words or stems for verbs or partial compound words.
Use fuzzy string match
Enables fuzzy string matches. (Default: Selected)
-
Select
OK to save the optimization settings.
The Optimization Settings window is closed.
-
Select
Start to start the optimization process.
During optimization, each parameter is modified and tested. Optimization ends when no further improvement can be reached, or if the classification result reaches 100%.
You can select
Stop at any time to terminate the optimization process. Any improvements
for the classification quality up to that point are saved.
-
When you are happy with the results, select
Close.
If auto optimization was able to improve classification quality, a message is displayed prompting you to save the changes to the project.
- If the classifier parameters were modified, you should retrain the project to apply the new parameters to the content classifier. On the Process menu, select the relevant training setting.
- On the Project tab, in the File group, select Save Project.
- Optional. Generate a classification benchmark and compare your new classification settings to previous settings.